MiniMax-M1 40k is an open-weight, large-scale reasoning model developed by MiniMax. It utilizes a hybrid Mixture-of-Experts (MoE) architecture integrated with a Lightning Attention mechanism to provide linear computational complexity for long-context tasks. The model is designed for advanced reasoning, particularly in mathematics, programming, and complex software engineering environments.\n\nThe model contains 456 billion total parameters, with 45.9 billion parameters activated per token during inference. The "40k" designation refers to its specific thinking budget, which represents the maximum number of output tokens allocated for reasoning chains. It was trained using a novel reinforcement learning algorithm called CISPO (Clips Importance Sampling weights), which improves training stability and convergence speed relative to traditional methods.\n\nMiniMax-M1 40k supports a native context window of 1,000,000 tokens, allowing it to process extensive codebases and lengthy documents in a single pass without segmentation. It has demonstrated high proficiency in benchmarks such as AIME 2024 and SWE-bench, highlighting its capabilities in logical reasoning and tool utilization.